LFM2.5-2.6B Matches DeepSeek-V4-Flash on Tool Calling; LEAP Fine-Tuning
Melvin Vivas · X post · 2026-08-08 · Open on X
Topics: AI Agents, Tool Use & MCP, Fine-tuning & Model Customization, Evaluation (Evals) & Testing · Level: intermediate
Summary
The creator is trying Liquid AI's LEAP framework for fine-tuning and says it looks like it has everything you need. He quotes a benchmark where Liquid AI's LFM2.5-2.6B was run against DeepSeek-V4-Flash on one machine with 4x RTX 5090. Each job only counted as complete if the model made every required tool call, and the small model reportedly matched DeepSeek-V4 on tool calling while running 3.7x faster.
Key points
- LFM2.5-2.6B is a small Liquid AI model that reportedly reaches DeepSeek-V4 level on tool calling.
- It reportedly ran 3.7x faster than DeepSeek-V4-Flash on the same hardware (4x RTX 5090).
- Test method: both models got the same three jobs, and a job only completed if the model made every tool call.
- Example job topics included a weather task.
- LEAP is Liquid AI's fine-tuning framework, which the creator says looks like it has everything you need.
Resources mentioned
- Liquid AI (@liquidai) on X · website · x.com · free
Liquid AI's family of efficient foundation models, used here as the example of a model you can run in the browser with ONNX.
Also in: Trying Liquid AI Decision Models (d1, d1 ultra) via API (Melvin Vivas on X · notes), Liquid RIPL: Low-Latency Inference Engine for LFMs (Melvin Vivas on X · notes), ONNX + Transformers.js: Run Models Like Liquid LFM in the Browser (Melvin Vivas on X · notes), Try Liquid AI's open d1-3B decision model in Colab (Melvin Vivas on X · notes) and 30 more - LEAP (Liquid AI) · tool · leap.liquid.ai · check price
Liquid AI's framework for fine-tuning and customizing its small models.
Also in: Smoke-Testing LFM Fine-tuning with Codex and Liquid AI's LEAP (Melvin Vivas on X · notes), Small Local Models + Fine-Tuning Instead of More Compute (Liquid AI, LEAP) (Melvin Vivas on X · notes) - LFM2.5-2.6B · tool · huggingface.co · free
Liquid AI's small language model, which can be paired with the LFM2.5-VL-3B vision model.
Also in: Coworker: Open-Source Work Agent That Runs on Small Local Models (Melvin Vivas on X · notes), Coworker with Liquid AI LFM2.5-2.6B via LM Studio on a Mac (Melvin Vivas on X · notes), Zero-Cost Coworker Setup: OpenRouter Free Models + Local LFM2.5 (Melvin Vivas on X · notes), Coworker: A Subscription-Free AI Agent on Local LFM2.5-2.6B (Melvin Vivas on X · notes) and 9 more - DeepSeek V4 Flash · tool · huggingface.co · free
Cheap DeepSeek model whose intelligence-index score is roughly equal to GPT-5's.
Also in: Is AI Getting Cheaper? How to Measure LLM Cost Per Task and Per Intelligence Level (Melvin Vivas on X · notes), Low-Cost Agent Run: DeepSeek V4 Flash via OpenRouter in ohmypi (Melvin Vivas on X · notes), Running Codex with DeepSeek V4 Flash through OpenRouter (Melvin Vivas on X · notes), DeepSeek V4 Flash at 90% Off on Nous Portal (Melvin Vivas on X · notes) and 4 more
Try this
- Check out Liquid AI's LEAP framework for fine-tuning.
- Try LFM2.5-2.6B for tool-calling tasks.
- Build a small tool-calling benchmark: give a small model and a large model the same multi-tool jobs, count a job as passed only if every required tool call is made, and compare pass rate and speed.
More in AI Agents, Tool Use & MCP
- Pi Coding Agent: Four Built-In Tools (read, bash, edit, write)
- Using a Hermes agent to watch the terminal and drive Codex
- Serve LFM2.5-2.6B with vLLM and connect it to Hermes
- Claude Code Agents Can Now Talk to Each Other Natively
- Prime Agent: A Self-Improving RLM Coding Harness with Code-Based Tool Calls
- Prime Agent: A Self-Improving RLM Coding Harness That Calls Tools in Code